Overcoming barriers to generalizability of deep learning for dental image analysis

• Applicant: Professor Dr. Falk Schwendicke
• Subject Area: Dentistry, Oral Surgery
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 561442884

While Deep Learning has shown immense potential across medical fields, relatively few research studies have successfully transitioned to clinical applications. Many promising models remain confined to research settings due to the difficulty in demonstrating consistent and generalizable performance across diverse patient populations. This is often caused by training on homogeneous datasets, which do not represent the complexities and variation that the model may encounter when tested on real-world data. In dentistry, according to our search, there were more that 1000 original studies involving DL since 2012. Despite that, only a handful of AI applications are currently available, failing to transfer the advantages that these models could bring into daily practice. Since generalizability is one of the main causes, the objective of the planned investigation is to provide a deeper understanding of how dental researchers should compile suitable datasets and set up their model training to achieve models that generalize well across data from various sources. This will enhance the performance of AI models and support their integration into the clinical practice. Furthermore, the outcomes of this project could be helpful in overcoming similar issues in other medical disciplines. The aims of the proposed work packages (WPs) are presented below: • WP1 – Demonstration and quantification of missing generalizability: The objective of the first package is to quantify generalizability by training and testing low hundreds of models on various combinations of data for selected tasks. These will include the detection and/or segmentation of apical lesions in panoramic radiographs, caries in bitewings and oral lesions in photographs. While being purely descriptive, this exhaustive evaluation will serve as a base for the following packages. • WP2 – Roots of missing generalizability: The second package aims to examine characteristics of dental imagery that are responsible for the missing generalizability. This will be assessed by splitting the available datasets into groups according to the meta-data, using explainable AI to identify features typical for data from different institutions, projecting data characteristics into a lower-dimensional space, and predicting model performance based on the similarity of data distribution. • WP3 – Achieving generalizability: The purpose of the third package is to implement the findings of WP2 and investigate methods that may improve generalizability. These include different learning approaches and the use of transformer models. • WP4 – Enhancing generalizability using synthetic data: The use of synthetic data may allow for standardization of training and/or testing data while alleviating privacy and ethical concerns associated with the use of AI. The aim of the fourth package is to produce synthetic data using generative models and assess their potential in improving model generalizability.

• DFG Programme: Research Grants